1
Ultra-Low Voltage Nano-ScaleEmbedded RAMsKiyoo Itoh, Hitachi Ltd.
IEEE SSCS DLP, Colorado, June 15, 2006
OUTLINE1. Introduction2. Trends & Challenges3. Low-Voltage Limitation of RAMs4. Leakage & Speed Variation of
Peripheral Logic Circuits 5. Memory Cell Size6. Future Prospects7. Conclusion
K. Itoh, Hitachi
2
Trends in RAM Developments (R&D)M
emor
y C
apac
ity/
chip
(bi
ts)
Year
256 K
1970
DRAM
first presentation atISSCC/Symp. VLSI Circuits
1980 1990 2000
16 G
1 G
64 M
4 M
16 K
1 K
SRAM
Year
Mem
ory
Cel
l Are
a (μ
m2 )
19901970 1980 2000
10,000
1,000
100
10
1
0.1
full CMOS
DRAM
SRAMTFT load
poly-Si load
full CMOS
3-D capacitor
ISSCC, Symp. VLSI Circuits
0.01
TFT
TFT
Planar capacitor
K. Itoh, Hitachi
3
Challenges to LV e-RAMs
RAM Cells•Extend low-voltage limitation to sub-1 V-Degraded S/N-Increased leakage
•Reduce cell size
K. Itoh, VLSI Memory Chip Design, Springer-Verlag, March 2001
e-RAM
RAM cellarray
DRAMSRAM
periph.DL
SRAMDRAM
WL VDD
0
"1"
"0"
DLWL
0 VDD
DL
Cs
Peripheral Circuits •Reduce leakage-Increased ISTB & IACT
•Reduce speed variation-Unreliable operations
K. Itoh, Hitachi
4
1. It is governed by soft-error of cells, or S/N of cells and cell-relevant circuits.
2. As long as ECC is used, it is governed by S/N.
3. S/N is determined by •Signal charge & signal voltage of cells, •Flip-flop circuits that DRAMs use for sense
amps, while SRAMs use for cells themselves.
Low-Voltage Limitation of RAMs
ECC: Error Checking and Correcting circuit
K. Itoh, Hitachi
5
Comparisons of Flip-Flop Circuits
SRAM CellDRAM SA
Standby
Active
Margin
Off with all nodes at VDD/2 On with static of all MOSTs
Dynamic sensing of vS i by ratio of M1 & M5
(SN on, then SP on) Static of other MOSTs
Sensitive to VT & ΔVT of Sensitive to VT & ΔVT ofonly two MOSTs, M1 & M2 all MOSTs → Narrow margin
CircuitDLDL
VDD
M1
M5
VDDi
VDD
0DL DL
VDD/2
VDD/2SN
SP
M1 M2
VDD/2 VDD/2
STB ACT
-vS
K. Itoh, Hitachi
6
Signal Charge QS of RAM CellsQS ≅ Soft-Error Qcrt . The larger the QS , the smaller the SER.QS decreases with device and voltage scaling.
QS = CSVDD/2 QS = CSVDD , CS =(C1 + 2C2 )CS ; Intentionally added, C1, C2 ; parasitic, small, andlarge, and needs to be rapidly decrease with devicegradually decreased scaling.with device scaling for SER is always larger than formaintaining large vSIG. DRAM.
WLVDDDL
0Cs
"1"
"0"
DRAM SRAMVDD
DLC1
DL
C2WL
0VDD
K. Itoh, Hitachi
7
Y. Nakagome et al., IBM J. R&D, Vol.47, No.5/6, Sep./Nov. 2003E. Ibe, The Svedberg Laboratory Workshop on Applied Physics,Uppsala, May 3, 2001
Qs reduced with capacitydue to VDD & device scalingSmaller QS of SRAM cell
SER depends on QS
DRAM; decreases with memorycapacity due to large intentionally-added CS & spatial scaling thatreduces charge collection.SRAM; increases with memorycapacity due to rapidly-decreasingparasitic CS despite spatial scaling.Solutions:•Increase in CS (SRAM cells)•Uses of triple well, redundancy,
ECC etc.
SER
Cro
ss S
ecti
on/c
hip
(cm
2)
Memory Capacity (bits)100K 1M 10M 100M 1G
1E-5
1E-6
1E-7
1E-8
1E-9
1E-10
DRAM
SRAM
1
10
100
1000
Sign
al C
har
ge, Q
S(f
C)
Memory Capacity (bits)1G64M4M256K16K
4
DRAM
SRAM
Trench
Stack
Planar
12.56.95.1
Stand-aloneQS =CS VDD/2
ΔCS(α-SER: x10-3.5)
Signal Charge (QS) of RAM Cells
K. Itoh, Hitachi
8
Error Checking & Correcting (ECC)
M. Horiguchi et al., IEEE J. SSC, 23, p. 27, Feb. 1988
ECC word = 128 data bits + 8 check bits, FIT = 10-9/hour
102 106 108104
SER without ECC (FIT)
SER
wit
h E
CC
(FI
T)
10-18
10-16
10-14
10-12
10-10
10-8
10-6
10-4
10-2
102
104
106
108
without ECC
periodic correction
(1 ECC word/7.8 μs)no corre
ction during 10-year period64 Mb
256 Mb
16 Mb
100
100
one upset/ 1 k hours
K. Itoh, Hitachi
9
Minimum VDD (Vmin) of RAMsDRAM SRAM
δVT : VT-mismatch between paired MOSTs, VT0 : Average VT
If only cross-coupled nMOSTs determine the voltage margin during read,VG = VDD –VT0 –δVT ≥ 0, ∴Vmin = VT0 +δVT
VDD
vS > δVT,vS ≅ (VDD/2)CS/CD
∴Vmin = 2δVT CD/CS
=10δVT (CD/CS ≅ 5)
VG =VDD/2 – (VT0
+δVT) > 0 ∴Vmin =2 (VT0 +δVT)
Cell SA Cell VDD
VT0 +δVT
DL DL
VT0
DL DLVDD/2−vS VDD/2
VT0 +δVT VT0
0
CD
DL
Cs−vS
WL
K. Itoh, Hitachi
10
Lowest Necessary VT0 for SA
subthresholdcurrent of M2
(isub)
VDD/2
WLon
SPon
–vS
SNon
DL
DL
0
C
A
B
δVTp
tS
Δv
DL
SP
SN
VDD/2
DL
VDD/2
0
VDD
M1 M2
–vS VTp
• Signal (-vS) is amplified, so DL isdischarged to A. After that, DL isgradually discharged by isub (M2).
• SP is on before DL reaches C.• VT0 (nMOST) ≥ 0.2 V @extra., 25°C,
if tS = 5 ns, Δv = 50 mV, CD =100 fF@ 120°C
K. Itoh, Hitachi
11
•Cross-coupled MOSTs need a high VT to ensure a small retention current through reducing iL.
•VT0 is the average in a chip, because it is the average that determines retention current of the chip.
•VT0 must be quite high and unscalable.
Lowest Necessary VT0 for SRAM Cell
Extrapolated VT0 (V) at 25 ºC -0.2 0 0.2 0.4 0.6 0.8 1.0
100 Lg =0.1 μmW (QT)=0.20 μm W (QD)=0.28 μm W (QL)=0.18 μm
Tj =125 °C100 °C75 °C50 °C
25 °C
high speed(0.49)
low power(0.71)10 μA
0.1 μA
10-2
10-4
10-6
10-8
1-M
b ar
ray
rete
nti
on c
urr
ent
(A)
Extrapolated VT =VT (nA/μm)+0.3 V
K. Itoh et al., CICC 2004 Dig. Tech. Papers, pp. 339-344, Oct. 2004
iL
VT0 +δVTVDD –VT0 –δVT
VDD
DL
K. Itoh, Hitachi
12
VT -Mismatch (δVT)
90 nm bulk
65 nm bulk
45 nm bulk
32 nm bulk
65 nm FD-SOI
45 nm FD-SOI
32 nm FD-SOI
Standard deviation (a.u.)210
σ (VT)σint
σext
σint
σext
σ (VT)
M. Yamaoka et al., Symp. VLSI Circuits 2004
1.VT Variation (ΔVT) as source of δVT
•Extrinsic ΔVT due to implant non-uniformities & Δ (L,W)
•Intrinsic ΔVT due to random microscopic fluctuations of dopant atoms in the channel area.
2.ΔVT & δVT increase with reducing MOST size evenfor a fixed generation.
σint ∝ 1/√LWσ(δVT) ≅ √2 σint
3.ΔVT & δVT increase withdevice scaling.
ΔVT in a chip has no room in time & area to be compensated for.
K. Itoh, Hitachi
13
Larger δVT of SRAM cell
SRAM Cell
DLDL
DRAM SA
Circuit
MOS Size 10 – 20 F 2 1.8 – 2.8 F 2
CircuitCount
in a chip
M/(64 -1024) M
δVT Small Large
64-1024
Relaxed SA layout
DL DL
M
Array
M : memory capacity
K. Itoh, Hitachi
14
Maximum δVT in a chipr : repairable percentage
SRAM Cell
90 65 45 32
400
0
Max
. δV
T (m
V)
100
200
300
10%
1%
0.1%0.01%0.001%
r = 0
F (nm)SRAM 32Mb 64Mb 128Mb 256Mb
LW = 2F 2, tOX = 1.9 nm
DRAM SA
90 65 45 32F (nm)
40
80
120
0
Max
. δV
T (m
V)
20
60
100
10%
1%
0.1%0.01%0.001%
r = 0
512Mb 1Gb 2Gb 4GbDRAM
LW = 20F 2, tOX = 1.9 nm64cells/SA
100 330
K. Itoh, Hitachi
15
Vmin (r = 0.1 %)
DRAM SRAM
*Actual Vmin determined by all MOSTs in a cell.
FD-SOI
1.4
1.2
1.0
0.8
0.6
0.4
0.2
090 65 45 32
512Mb 1Gb 2Gb 4GbF (nm)
Vm
in(V
)
Cell
SA
LW = 20F 2, tOX = 1.9 nm 64 cells/SA, VDD/2 DL pre.VT0 = 0.2 V
0.6 FD-SOI
1.4
1.2
1.0
0.8
0.6
0.4
0.2
090 65 45 32
32Mb 64Mb128Mb256MbF (nm)
Vm
in(V
)
LW = 2F 2, tOX = 1.9 nmVT0 = 0.49 V (HS), 0.71 V (LP)
LP
HS
0.9
1.1HS*
K. Itoh, Hitachi
16
Approaches to LV SRAMs
1. Use ECC & Redundancy.2. Minimize ΔVT & δVT.
•Large cells with largeMOSTs despite losingbit density
•Symmetric cell layout3. Stay at a high VDD (≥1 V)
due to its still large ΔVT& δVT of bulk CMOS.Even so, power-supply control cells needed for small subthreshold current.
4. Extend low-VDD limitation to sub-1-V with FD-SOI.
M. Khellah, et al., ISSCC Dig., pp. 624-625, 2006
13612311295
1.2
1.0
0.8
0.6
0.4
0.2
0
activestandby
Bit Density (Mb/cm2)
Vm
in(V
)
K. Itoh, Hitachi
17
Sources of extrinsic δVT in the conventional cell:•Pattern deformation after processing •Mask misalignment •Local size fluctuationSolution: Lithographical symmetric cell (“Thin” Cell)
K. Osada et al., IEEE J. SSC, vol. 36, No. 11, pp. 1738-1744, Nov. 2001.M. Kanda et al., Symp. VLSI Tech. Dig. Tech. Papers, pp. 13-14, June 2003.F. Arnaud et al., Symp. VLSI Tech. Dig. Tech. Papers, pp. 65-66, June 2003.
Dotted area: after processingdiffused
poly gate
Conv. (2 cells)
M1 M2
LS cell (2 cells)
M1M2
M1 M2
Symmetric Layout for Small δVT
•Reduced δVT by simple patterns suitable for OPC •DLs shielded by power lines
OPC: optical proximity correction
K. Itoh, Hitachi
18
Power-Supply Control Cellsfor small subthreshold currents
K. Itoh, ICICDT2005 Dig.
DL
VDD+δVD
DLVDD (0) 0 (VDD)
• High VT to reduce iL• δVD to offset a high VT
& δVT
Low leak, wide margin & low power with low-VDD DL.Unscalable MOSTs needed.
Boosted Power Supply Source-Line Driving
• Raised source duringSTB to reduce iL withincreased VT of off-MOST
Reduced margin during STB by δVS (>0.3 V)
VDD
ACT δVSSTB
0
source
K. Itoh, Hitachi
19
Source-Line DrivingAlong with reduced DL voltage at active-standby transition
K. Osada et al. ISSCC2003 Dig. pp. 302-303
Electric-field relaxation90% reduction ingate leakage & GIDL
Sub-S backbias90% reduction insubthreshold leakage
G-S backbias100% reduction insubthreshold leakage
Standby
Active
DL0.5 V 1.0 V1.0 V0.0 V 1.5 V1.5 V
DL VSS
1.5 V
WL 0 V
0.5 V 1.5 V
K. Itoh, Hitachi
20
Measured Retention Current of Cell
K. Osada et al. ISSCC2003 Dig. pp. 302-303
Successful Application1.5-V 27-ns 6.42 x 8.76 mm2 16-Mbusing ECC with 3.2-ns/9.7% speed/area penalties.
(1) Leakage still large1.6 μA for 16 Mb despite high VT , thick tox, and S-driving.(2) Reduced QS in standby modeThe cell power-supply decreases by the raised source voltage.Further low-VDD operation may behazardous, even if ECC is used.
Limitations and Challenges
Conv.
Prop.
25ºC
Sub. + GIDL 48.5
PMOSNMOS NMOS
95 fA
PMOS
17 fA
Tunnel 46.5
3 14VT (extrap.) = 0.7 V(N), -1 V(P)tox (electrical) = 3.7 nm
NMOSConv.
102 fA
PMOS
1244 fA
Prop.
PMOSNMOS
90ºC
Sub. + GIDL 1182
Tunnel 62
81Subthreshold currentsensitive to temp.
K. Itoh, Hitachi
21
Source-Line Driving
=VDD –Vmin(STB)
VDD
Vmin(STB) SRAMarray
δVS
+
VREF
ACT STB
M. Khellah, et al.,ISSCC Dig. pp. 624-625, 2006
Vmin(STB) : Min. VDD to retain the data of all cells in the array.
Lower VT0 → Larger δVS
→ Higher VT0 with deeper body bias (ΔVT)→ Lower iL
to reduce leakage & its variation while retaining the data
VD
istr
ibu
tion
VDDVmin(STB)
VT0
ΔVT δVS
VTmax
K. Itoh, Hitachi
22
Double-Gate FD-SOI•Small ΔVT & negligible δVT
(ultra-thin & lightly-doped channel)•Adjustable VT → multi-VT
•Large VT change(wide-range well-bias control)
•Reduced SER & small ipn
•Dynamic VT MOST(e.g., G-well connection)
M. Yamaoka et al., Symp. VLSI Circuits 2004, R. Tsuchiya et al., IEDM2004 Dig. pp. 631-634
Back-gate voltage Vwell (V)
0.2
0.1
0.0
-0.1Thre
shol
d vo
ltag
e V
t(V
)
-1.0 -0.5 0.0 0.5 1.0
nMOSTVt @VDS =1.0 VtSOI =20nm
tBOX = 100 nm
10 nm
B
A
σint90 nm bulk
65 nm bulk
45 nm bulk32 nm bulk
65 nm FD-SOI45 nm FD-SOI
32 nm FD-SOI
Standard deviation (a.u.)210
σ(VT)σext
σintσext σ(VT)
σint
G
S D
p+(VBN )
G
S D
n+(VBP )
p+
S DG
S DG
n+
n-well
p-sub
NMOST
SOI( ≤ 20nm)
BOX (≤ 10nm)
PMOST
p-sub
K. Itoh, Hitachi
23
M. Yamaoka et al., SOI Conf. Dig. pp. 109-111, Oct. 2004
SRAM Cells with Dynamic-VT MOSTs
M. Yamaoka et al., A-SSCC Dig. pp. 109-112, Nov. 2005
M1: decreased VT M2: increased VT
M3: increased VT M4: decreased VT
DL DL
VDD0VDD
VDD
M1 M2
M3 M40
Write margin improved withdecreased VT for driver/transferMOSTs & increased VT for load MOSTs.
W
0
DL DL
VDD
0 R1.0 V
RW 1.0 V
M1 M2
M3 M4M5 M6
to widen the voltage margin
K. Itoh, Hitachi
24
Challenges to LV e-RAMs
RAM Cells•Extend low-voltage limitation to sub-1 V-Degraded S/N-Increased leakage
•Reduce cell size
K. Itoh, VLSI Memory Chip Design, Springer-Verlag, March 2001
e-RAM
RAM cellarray
DRAMSRAM
periph.DL
SRAMDRAM
WL VDD
0
"1"
"0"
DLWL
0 VDD
DL
Cs
Peripheral Circuits•Reduce leakage-Increased ISTB & IACT
•Reduce speed variation-Unreliable operations
K. Itoh, Hitachi
25
Gate tunneling current (iG)• Insensitive to VG & temp. • Sensitive to tox
1/10 iG -reduction with tox -increment of only 2-3 Å for SiO2, whilethe same reduction with VG -decrement of as much as 0.5 V.Such a large VG control in low-VDD region is risky.
Device designers are responsible for the reduction. (High-k)
Subthreshold current (iL)• Insensitive to device structures• Sensitive to VG ,VT & temp thatcan be controlled by circuits.
1/10 iL-reduction with VT -increment, orVG -decrement of only 100 mV.
Circuit designers are responsible forthe reduction.
Leakage Currents of Periphery
iL ∝ W 10(VG-VT )/S
S ~ 100mV/dec.@100ºC 1/10 with ΔVT = 100mV
or ΔVG = -100mV
VG
iLW
K. Itoh, Hitachi
26
Subthreshold Current (iL) of Periphery
K. Itoh, VLSI Memory Chip Design, Springer-Verlag, March 2001Y. Nakagome et al., IBM J. R&D, Vol.47, No.5/6, Sep./Nov. 2003
1. Input-Predictable LogicDesigners can preparethe schemes in advance.
2. Slow Cycle (tRC =25, 60 ns)Each circuit is active for onlya short period within “long”cycle, enabling additional time for iL-control.
3. Iterative-Circuit BlocksMajor iL sources. All circuits in each block areinactive, except selected one.
4. Robust CircuitsiL-immune NAND dec. (w/o iL-sensitive NOR dec.)
5. On-Chip Power SuppliesVDH &VBB utilized for dual-static VT.
non
-se
lect
edsele
cted selected WL
sele
cted
YL
Rowdec.
non
-se
lect
ed
WD Memory array
selected
non-selected
non-selected
Col.dec.
φP
axiVDH 0φX
dynamicNAND
ayi
VDD 0
φY
staticNAND
CLK
axi
WL
selected
non-selectedVDH
tRC
ayi
YLVDD
VDH
VDD
Features of RAM Periphery
K. Itoh, Hitachi
27
•At present, VT is still so high that iL is small in active mode, though iL dominates in standby mode.
•In the future, with further reducing VT, iL will dominate evenin active mode. Leakage reduction for active mode is the key.
Cu
rren
t (A
)
Trends in DRAM Peripheral Current
Capacity (bits)16M 64M 256M 1G 4G 16G 64G10-6
10-5
10-4
10-3
10-2
10-1
100
1011.2 A
IAC
IACT
IDC (IL)
Cycle time=180 nsT = 75°C, S = 97 mV/dec.
0.160.24 0.190.53 0.40 0.32 0.13
VDD (V)
Extrapolated VT at 25°C (V)
1.01.21.53.3 2.5 2.0 0.8
K. Itoh, VLSI Memory Chip Design, Springer-Verlag, March 2001
iL-Increase in Periphery
IACT = IAC +IDCIAC = ΣCj VDD fIDC = ΣiLk ∞ ΣWk 10-VT/S
VT = aVDDk >> j
chip (active)
Inactive (k)
Active (j)
K. Itoh, Hitachi
28
•Use high-VT MOSTachieved statically or dynamically with VTh =VTl + ΔVT
•For static high-VT
ΔVT by ion impla. orstatic-VBB application.
•For dynamic high-VT
ΔVT by dynamic back-biasing schemesG-S back-bias is best due to large ΔiL/Δδ, applicable even to active mode.
Offset Voltage, δ (V)
Leak
age
Red
uct
ion
(ra
tio) 1
10-1
10-2
10-3
10-4
0 0.2 0.4 0.6 0.8 1.010-5
L = 90 nm, tOX = 2 nmVDD = 1 VS = 100 mV/decadeK = 0.2 V1/2, 2ψ = 0.6 Vλ = 0.05
VDD
0
+δ0
VDD
0
0
-δ
VDD
0
0
-δ
VDD
+δ0
VDD -δ0
0
0(DIBL)
NMOS
∼−ΔVT
k (√δ +2ψ -√2ψ )ΔVT δ∼−
Basic Concept of iL-Reduction
Y. Nakagome et al., IBM J. R&D, Vol.47, No.5/6, Sep./Nov. 2003
K. Itoh, Hitachi
29
Three Practical Reduction CircuitsApplicable even to Active Mode
Switched-Source Imp.(G-S Self-Backbiasing)
Power Switch utilizinginternal power supply(G-S Offset Driving)
Dual-Static VT utilizing internal power supply
(VDD)
IN
(0)
VDD
OUT
SSI
VDD
OUT
(0)(VDD)
IN
SSI
VDD
VDD
low VT–VBB
low-VTcore
–VBB
VDD VDD
VDH
(>VDD)
VDDVDH (>VDD)
low VT0
low-VTcore
δ
VDD - δ
K. Itoh, Hitachi
30
No matter how large i1 is, itis confined to const. current i2’ with self-adjusting δ.
Stacking effects (δ = 0.2 V)• G-S backbias of Q1 (1/100)• Sub-S backbias of Q1 (1/1.5)• DIBL effect of Q2 (1/2)
Applicable even to active mode• Fast iL-control capability with
small δ & CL and self-reduction• Small area penalty if applied
to iterative circuit blocks• Capability of confining to min.
active circuitry
SSI (G-S Self-Backbias)
M. Horiguchi et al., Symp. VLSI Circuits Dig., p. 47, 1993.
for fast iL-control of input predictable logic
i1’ = i2’∴δ = (S/ln10) ln(W1/W2)
Reduction Ratio γ =i1’/i1 =10−δ /S =W2/W1
• Smaller iL (=i2’) & larger δ with smallerW2
• γ = 1 (no reduction) for W2 =W1Other secondary effects reduce iL.(Sub-S backbias & DIBL)
circuits
Q1
0
i0 : current density
i1=i0W110-VT/S
w/o SSI with SSI
Q1
0
Q2
0
const. current
δ
i2’=i0W210-VT /S
i1’ =i0W110- (δ +VT)/S
SSI
W1 W1
W2
K. Itoh, Hitachi
31
Inverter Chain
• For 0-V input, iL flows from an n-MOST in each 0-V input inverter, and accumulatesinto SSI.
• SSI confines to its constantcurrent (=i0 W2 10-VT/S ).
•W2 ≅w1 ≅ ··· ≅wn without speedpenalty because each inverterswitches at different timing.
• Area penalty is negligible with increasing the number ofinverters because W2 « Σwi .
Such is the case for p-MOST SSI.
Small Area Penalty with SSI Sharing
M. Horiguchi et al., IEEE J. Solid-State Circuits, vol. 28, no. 11, pp. 1131-1135, Nov. 1993.
ΦDVDD
VDD
ΦS0
VDD
0VDDIN
0
SSI
SSIVDD -δ
δ
w1
W2
w20
K. Itoh, Hitachi
32
• During non-selected periods iL (= i’ ) flows from each circuit, and accumulates into SSI.
• SSI confines to its constant current (=i0 W2 10-VT/S).
• W2 ≅w without speed penalty because only one MOST is activated with SSI on.
• Area penalty is negligible with n » 1 because W2 « nw.
Small Area Penalty with SSI Sharing
M. Horiguchi et al., IEEE J. Solid-State Circuits, vol. 28, no. 11, pp. 1131-1135, Nov. 1993.
Iterative Circuit Block (non-selected/standby)
= 0Q1
W1 =nw
i1’=ni ’
i ’0wQ
0Q2
SSI
i ’0wQ
W2
n
δ
K. Itoh, Hitachi
33
Confining to Minimum Active CircuitryPartial activation of multi-divided block
K. Itoh, VLSI Memory Chip Design, Springer-Verlag, March 2001
1,023i → 134ifor n/m =1024/8
block (selected)
n
(n -1)i
activatedQ 0
i0Q VDD
i0Q VDD
VDD w
i =i0 w 10-VT/S
subblock subblocks(selected) (non-selected)
i2’=i0 W2 10-VT/S ≅ i , W2 ≅ w
n/m
#0 #1 #m -1
(n/m -1)i
i0Q
VDD
SSI
VDD
Q
0
0Q2
δ
• • •
0
δ
0
VDD
≅ (m -1)i
W2 W2
w
i2’ i2’
K. Itoh, Hitachi
34
Low-VT switch (Q) shuts off the supply of low-VT core during standby. A raised VDH needed to cut off Q with G-S backbias.
Problems:1. If VDH generated by charge pump, • Unregulated floating VDH
For well-regulated VDH,C2VDHf < CPVDDfP , C1 »CP , C1 »C2 , level monitor.
• Increased pump powerFor low pump power with keeping the VDH level, CPVDDfP & C2VDHf reduced.-> Smaller C2 & slower f
2. Area penalty by large Q3. Slow recovery of internal power node ( )
Power Switch (G-S Offset Drive)
K. Itoh, VLSI Memory Chip Design, Springer-Verlag, March 2001
VDH Gen.C1 C2
φlow VT
fP VDD CP
VDD VDD
VDH (>VDD)
0
ACT STB
Low-VTcore
Q
φ
f
Levelmonitor
K. Itoh, Hitachi
35
Applicable even to active mode, if the switch itself operates fast enough. After evaluating the input, the output level continues to be held by high-VT holder without leakage. Otherwise
•Floating output discharges, causing a large iL at pMOST in the succeedingcircuit, in which the switch is still on.
•Unnecessary discharging prevents the output from quick recovery.
Power Switch with Level Holder
VDD
Level holder
in outVDD
Low-VTcircuit
VDD
Low-VTcircuit
High-VT
off
VDD with
w.o.
0
iL
K. Itoh, VLSI Memory Chip Design, Springer-Verlag, March 2001
K. Itoh, Hitachi
36
Dual Static VT
K. Itoh, VLSI Memory Chip Design, Springer-Verlag, March 2001
iL reduced to 1/5 for uniform use of VTlwith assumptions: Wtotal (critical path)=10% of Wtotal (chip), VTl = 0.21 V, VTh = 0.31 V, S = 0.1 V/dec.
highVTlowVT
criticalpath Chip
Selective use of a high VT to• Off-MOSTs during standby• Non-critical path, while using
low VT to critical path.-> Low iL & high speed chipThe reduction is not remarkablebecause VT difference must besmall. A large VT difference may cause a racing problem: a pulse-timing imbalance between VTl -& VTh-circuits.
Useful for active & standby modesVTh
VDD
VTl
'L''H''L'(standby)
VDD
VTl
VTh
K. Itoh, Hitachi
37
256-Mb DRAM (Standby)
VDH
(VDH)
VDD
VDH
Memory array
VDH
RX
I/O SA
Sub-word driver
SWLVDH
VBB
MWLVDH
VBB
VDHL
φH
VDDL
(VDH)
(VDH)(0)
(0)
(VDH)
YL(0)DL DL
VBB
φD
(VDH)RX
VDHArray Control
(VDH)
(VDD)SSI2
SSI1
fromrow dec.
Column decoder & driver
Sense amplifier
Main-word driverM. Hasegawa et al., ISSCC Dig. pp. 80-81, 1998.
K. Itoh, Hitachi
38
M. Hasegawa et al., ISSCC Dig. pp. 80-81, 1998.
40 µA with dual VT
Conv.
ProposedPower downself-refresh
1.28 mA
0.21 mA
Dual static VTSSI
T = 75°CS = 109 mV/dec. (PMOST)
89 mV/dec. (NMOST)VT = 0.03 V (PMOST, 0.7 nA/μm)
-0.02 V (NMOST, 0.7 nA/μm )VDD = 2.0 V, VDH = 3.8 V, VBB =-1 V
Peri. Drivers Y-dec Arraycontrol
Standby Current Reduction (256 Mb)
K. Itoh, Hitachi
39
Conv.
1180 mA
Proposed
69
SA
VDH = 1.75 V, VDD = 1 V, tRC = 180 nsVT = -0.12 V (2nA/μm), S = 97 mV/dec., 75°C
AC
75
DC
word drivers695
(subthreshold) 1105
decoder209
others132
drv.
75
116
41
SSI power switch with level holder
1-V 16-Gb DRAM
T. Sakata et al., 1993 Symp. VLSI Circuits.
Active Current Reduction
K. Itoh, Hitachi
40
K. Hardee et. al, ISSCC2004 Dig. p. 200
0.6-V 16-Mb e-DRAM
• Some int. voltages are controlled to compensate for Δ (iL, τ) caused byΔ (VT, T, VDD).• In sleep mode, 0.3-VG-S b.b. reduces SA’s iL.
0.9 V
VNB
VNSB
on-chip converters
VPB
VPSB
3 V
ExternalVDD (0.6 V) VDD I/O (2.5 V)
-0.3 V
ND
0.9
-0.3
0.9
PD
VSPVDD
VPBSSI2PG
VSN
SENVPB
VNB
VDD
VNB SSI1
0
0.6 V-δ
0.6 0
-0.3
δNG
P-switch driver (inverter chain)
0.6
K. Itoh et al., CICC 2004 Dig. Tech. Papers, pp. 339-344, Oct. 2004
Y Circuits 0.6 V
-0.3
VPSB
DL0.3 V
VDD
VPB
VNSB
VNB
DL
40 fF
SA0.2-VVT
ND
PD
DL0
256WLs
0.60.9
DL
P-switch
P-switch
K. Itoh, Hitachi
41
2.5 V 3 V
0.6-V 16-Mb e-DRAM
2.5
SSI2XE1 0
2.5
3
2.5
PRE
ai
dec.
VNB
XE3 0 SSI3
3
level shifter
word driverXE2
0
3
XDE SSI1
VNB
33
WL0
WLn-1
2.5
0
0
2.5
0
3
0
2.5 K. Hardee et. al, ISSCC2004 Dig. p. 200K. Itoh et al., CICC 2004 Dig. Tech. Papers, pp. 339-344, Oct. 2004
X Circuits (Active)
K. Itoh, Hitachi
42
X Circuits (Sleep)
0.6-V 16-Mb e-DRAM
XDE-0.3
32.5
2.5 V 3 V
2.5
SSI2XE1 2.5
2.5
3
2.5
PRE0
ai0
dec.
VNB
XE3 3 SSI3
3
level shifter
word driverXE2
2.5
3
SSI1
δ1
3 -δ3
3
VNB
33
2.5-δ2
WL0
WLn-1
0
0
0
0
K. Hardee et. al, ISSCC2004 Dig. p. 200K. Itoh et al., CICC 2004 Dig. Tech. Papers, pp. 339-344, Oct. 2004
K. Itoh, Hitachi
43
Active Mode4-bank arch. with one-bankactivation confines active circuitry to 1/4, and reduces•AC power of cont. signals•iL in inactive banks if SSI
is applied to WD & cells.SSI1 : Small δ1 & CL → 0.3 nsSSI2 : Drawbacks;
Large δ2 (0.4 V) & CL → 3 nsCell-supply reduced by δ2
Sleep ModePeriphery off withpower switch off → 3ns
Multi-Bank Architecture
M. Yamaoka et. al., ISSCC2004 Dig. p. 494K. Itoh et al., CICC 2004 Dig. Tech. Papers, pp. 339-344, Oct. 2004
1.2-V 1-Mb e-SRAM
VDD
BS0
δ2SSI2
SSI1
0
other periphery(SA, write amp, etc.)
BK0BK1
BK2BK3BS3
BS1
BS2
VDD
BS0
VDD
WL0
0
0
WL1
WLn-1
VDD
0δ2 δ2
WD0
VDD -δ1
K. Itoh, Hitachi
44
Leakage of 1-Mb e-SRAM (Active)
M. Yamaoka et. al., ISSCC2004 Dig. p. 494
VT = 0.4/0.3 V
cells (VT = 0.4 V) word drivers amp.Conv.
(300 MHz)
SSI1(300 MHz)
SSI1 & SSI2P-switch
(100 MHz)
460 μA(1)
350 μA(0.75)
150 μA(0.33)
1.2 V, room temp.
K. Itoh, Hitachi
45
Challenges to LV e-RAMs
RAM Cells•Extend low-voltage limitation to sub-1 V-Degraded S/N-Increased leakage
•Reduce cell size
K. Itoh, VLSI Memory Chip Design, Springer-Verlag, March 2001
e-RAM
RAM cellarray
DRAMSRAM
periph.DL
SRAMDRAM
WL VDD
0
"1"
"0"
DLWL
0 VDD
DL
Cs
Peripheral Circuits •Reduce leakage-Increased ISTB & IACT
•Reduce speed variation-Unreliable operations
K. Itoh, Hitachi
46
Ever-increasing ΔVT,and rapidly-lowering gate-over drive with device scaling.They enhance speed variation of peripheryΔτ/τ ∝ ΔVT/(VDD –VT)Solutions
•For inter-die Δτ,Compensation with VBB generator. e.g., Speed improvement by 63%
•For intra-die Δτ, FD-SOI
Speed Variation90 nm bulk
65 nm bulk
45 nm bulk
32 nm bulk
65 nm FD-SOI
45 nm FD-SOI
32 nm FD-SOI
Standard deviation (a.u.)210
σ(VT)σintσext
σintσextσ(VT)
Process Technology (nm)
4
3
2
1
0 600 32250 130 65LP SOC
MPUITRS 2001
Gat
e O
ver
Dri
ve ,
VD
D –
VT
(V)
M. Yamaoka et al., Symp. VLSI Circuits 2004
K. Itoh, Hitachi
47
IDS (M1) is a good indicator of iL& speed. VGS (M1) = VDD/2 ≅ VT.VD is compared to VDD/2 + Δ andVDD/2 – Δ to determine if VNBshould be increased or decreased.
e.g.,For low VT (fast process or high Tj)VD < VDD/2 – Δ. The lower OP activates PUMP, so VNB starts to decrease and VTis increased to compensate. For high VT (slow process or low Tj)VD > VDD/2 + Δ. The upper OP discharges M2-gate for driving the body, allowing VTto be reduced and compensated for.Such is the case for VDD.
VBB Generator for NMOS Body
K. Hardee et. al, ISSCC2004 Dig. p. 200Temp (°C )
-25 0 25 50 75 100 125Fast Process
Slow Process0.6 V (VDD)
0.7 V
0.8 V
0.6 V
0.7 V0.8 V
0.4
0.2
0.0
-0.2
-0.4
-0.6
VN
B(V
)
0.7 V(VDD )
M12.5 V
M2
VD
VDD /2
VDD/2– Δ
2.5 V
VDD
VDD/2+ Δ
2.5 V
PUMP
VNBLIMIT
2.5 V
Max.=1/2VDD
Min.=-2/3VDD
φ
φ
K. Itoh, Hitachi
48
Intra-Die Speed VariationLow-Power CMOS LSIs
0.9 0.8 0.7 0.6
0
1
2
3
4
90 65 45 32
τ(V
T0 +
ΔVT)
/τ
(VT0
)
bulk SOIΔVT (+)
ΔVT (–)
1 1.29 1.89 2.51– 0.23 0.38 0.50
VDD(V)F (nm)
σ(VT) bulk (ratio)σ(VT) SOI (ratio)
L/W = 1F/6FVT0 = 0.3 V, ΔVT = 3σ(VT)τ (VT) ∝ VDD/(VDD –VT)1.25
0.530.861.19
3.76
0.861.19
(ITRS’03)
+–
K. Itoh, Hitachi
49
Challenges to LV e-RAMs
RAM Cells•Extend low-voltage limitation to sub-1 V-Degraded S/N-Increased leakage
•Reduce cell size
K. Itoh, VLSI Memory Chip Design, Springer-Verlag, March 2001
e-RAM
RAM cellarray
DRAMSRAM
periph.DL
SRAMDRAM
WL VDD
0
"1"
"0"
DLWL
0 VDD
DL
Cs
Peripheral Circuits •Reduce leakage-Increased ISTB & IACT
•Reduce speed variation-Unreliable operations
K. Itoh, Hitachi
50
RAM Cells (DRAM)Short DL allows a small CS & simple CS-structure with small CD.Planar-CS cell might replace e-SRAM cells. vsig ≅ CS/CD ⋅VDD/2.In addition, short DL enables low-VDD fast operation.
proposed (ISSCC2005) conventional
Cells/DL 32 128CS 5 fF (Ta2O5; MIM) ≥ 15 fF (MIS)
Additional wire No local wire (M0)Thermal budget no impact on logic intolerable impactCell RC delay W storage cont. Non-metalized cell
Co-salicided S/DCell contact R 10 Ω 10 kΩ
M. Iida et al., ISSCC2005 Dig. p. 460, M. Shirahata et al., ISSCC2005 Dig. p. 462
M3
M2
DL (M1)
Logic DRAM
M1
DL(M0)
M3
M2
M1 DL
Logic DRAM
K. Itoh, Hitachi
51
Stacked TFT SRAM Cells
S.-M. Jung, et al., Symp. VLSI Tech. June 2004, Y.H.Suh et al., ISSCC2005 Dig. p. 476.
Cell Size Reduction (6-T SRAM Cell)
• Single-crystal TFTThe highest density cell(25F 2) comparable to DRAM cells.1.8-V 61.1-mm2 144-MHz256-Mb SRAM.
• Drawbacks as e-SRAMsSophisticated process, High-VDD operation due to TFT PMOST ofS =140 mV/dec.,IDS = 2/3 of the bulk.
Load p-TFTs & transfer n-TFTsdouble-stacked over bulk driver n-MOSTs in different levels of layers.
K. Itoh, Hitachi
52
Cell-Size Comparisons
Cel
l siz
e (r
atio
)20
01-T 1-T 4-T 6-T
15
5
(Sta
nd-
alon
e 8
F 2
)
1
20.5
3-T 3-T
3.65.6
6.7
SAC for stand-alone DRAMNo SAC for others
1.8
10
3 polySAC 3 poly
3
L.I.
3D
3D
14
3
K. Itoh, Hitachi
53
6-T SRAM Cell: Due to high necessary VT & large ΔVT•Not suitable for sub-1-V VDD,•Continue to be used for a high VDD (≥ 1 V).Challenge; Small-ΔVT MOSTs. TFT cells for stand-alone SRAMs.
1-T DRAM Cell: •Suitable even for sub-1-V VDD. Challenges; Planar capacitors, Small-ΔVT MOSTs.
Peripheral Circuits:•Subthreshold-currents will be reduced sufficiently with existing techniques even for active mode.
•Speed variations will continue to be serious.Challenges; VBB control, Small-ΔVT MOSTs.
Two Approaches:High-VDD bulk-CMOS for low-cost RAMs,Low-VDD FD-SOI for high-speed low-power RAMs.
Future Prospects for RAMs
K. Itoh, Hitachi
54
1. I discussed challenges and trends in LV RAMs.2. I reviewed state-of-the-art LV RAM circuits.3. I gave prospects of RAMs with emphasis on
further needs for•Ultra-low voltage RAM cells, •Advanced devices & circuits to reduce speed
variations.
Conclusion